Study Of Student Class Behavior Recognition Using Deep Convolutional Neural Network

The recognition of students' classroom behavior using teaching video is one of the current research hotspots in the field of intelligent education.The statistics of the results of students' classroom behavior is beneficial to the real-time management and analysis of classroom teaching quality,and also provide a good research basis for improving teaching quality and determining the direction of education reform.This paper takes the classroom surveillance video of adult education as the research object,and makes a corresponding research on facial region detection,facial authentication and skeleton feature extraction based on convolutional neural network.The main work of this paper is as follows:First of all,we used the surveillance video to detect the faces of students,and the student's identity is verified according to the results.In this paper,the network of LFFD is cut and improved according to the actual classroom scene while balancing the accuracy of detection and the speed of calculation,and finally captured faces of smaller size,and achieved 98% facial detection accuracy and 32.4FPS detection speed in the test dataset.According to the results of facial detection,depth feature extraction and feature distance calculation are carried out to verify the identity information of the students in the class,and the accuracy of facial verification is 97%.The second part,we recognize students' behavior based on the graph structure.On the basis of facial detection,head posture detection results are introduced to provide auxiliary information for recognition.As a result,the recognition accuracy of "listening" behavior is improved by 12.02%.At the same time,in order to solve the problem of low accuracy of predicting missed body nodes,this paper proposes a method of generating student individual prediction box to assist predict.The results show that this method can achieve 98.48% accuracy of node prediction in the case of serious classroom occlusion.Finally,the spatial and temporal features of the dynamic skeleton are extracted based on the detection results of the students' upper body nodes,and the corresponding behavior prediction and classification are obtained.In the last part,for some behaviors with low recognition accuracy,this paper introduce hand object detection to construct different human-object association convolution networks.Compared with the recognition only by physical behavior,the accuracy of "writing" and "playing with mobile phone" behaviors increased by 24.42% and 25.86%,and reached 92.96% and 95.88% respectively.At the same time,for the problem of information frame redundancy in the process of making student behavior data set,a method of extracting key frames is proposed,so as to retain the information frames that contribute more to behavior recognition and reduce the redundancy of behavior data.